AI Agents Are "Raising Horses": A More Terrifying Revolution Than the Large Model Arms Race

Late at night on May 7, a large number of users discovered that DeepSeek's service was experiencing anomalies, and messages sent received no replies.

Soon "DeepSeek crashed" shot to the top of trending searches. Early the next morning, DeepSeek officially responded—service had been fully restored.

This is not the first time. ChatGPT has crashed, Claude has crashed, Kimi has crashed. The "downtime moments" of large models happen almost every month.

But this time, a stranger piece of news caught my attention—"AI horse-raising is here"。

What is "AI horse-raising"? It's not about having AI raise real horses, but rather a new type of "AI species" appearing on the internet—they are no longer simple chatbots, but can autonomously learn, autonomously decide, and autonomously actAI Agent. Just as horses need to be tamed, these Agents need to be "fed" data and scenarios; they will grow, make mistakes, and even "bolt".

Two days ago, an AI Agent, while operating a production database, accidentally deleted an entire table, causing a company's business to be interrupted for 3 hours.

This is a signal.

AI Agents are accelerating into our workflows, but at the same time,the "horse" is bolting.

One, from "tool" to "co-pilot" to "driver"

Over the past two years, AI's role has undergone three leaps.

First stage: tool. In 2023, ChatGPT burst onto the scene, and people used it to write emails, write copy, and look up information. AI was a hammer, and the person was the hand holding the hammer.

Second stage: co-pilot. In 2024, Copilot, Codex, and Claude Artifacts turned AI from a "tool" into a "partner". You write half, and it helps you continue; you build the framework, and it helps you fill in the content. The human role shifted from "operator" to "instructor".

Third stage: driver. In 2025-2026, AI Agents began to truly "drive". They no longer wait for you to give instructions, but autonomously complete tasks—scheduling meetings, managing projects, operating databases, writing code and deploying it live.

This is not a sci-fi movie. This is what is happening today.

Data shows that in the first quarter of 2026, global enterprise-level AI Agent deployments grew 470% year over year. Gartner predicts that by 2027, more than 60% of enterprises will use AI Agents to handle core business processes.

But the faster the growth, the greater the risk.

That Agent that accidentally deleted the database had the task of "cleaning up expired data". But what it understood as "expired" was not the same as what humans understand as "expired". It deleted a user table that was still in use.

This is not AI "breaking down". This is boundaries not being clearly drawn.

Two, low-code + AI: behind a 300% surge in efficiency

Among today's hot topics is one: low-code platforms + AI, software development efficiency increased by 300%.

How is this number calculated?

Traditionally, developing a medium-sized enterprise management system usually takes 3 developers, 3 months, and 300,000 lines of code. Now, with low-code platforms + AI Agents, one person, one week, and even without writing code, can build a prototype.

The efficiency improvement is not 300%; in some scenarios it is 1000%.

But problems come along with it.

The spread of low-code and AI has made the threshold for "building" extremely low. In the past, if you wanted to make a system, you had to know programming, understand architecture, and grasp databases. Now, you say to AI, "Help me build an order management system," and it builds it for you in a few minutes.

The lower the threshold, the more uneven the quality.

I have seen a case: a company used AI Agent + low-code to build an internal approval system in three days, and after launch everything was normal. Until a month later, an audit found that a 500,000 reimbursement had inexplicably passed approval—the Agent had "misunderstood" the approval authority when handling the process.

It was not that the Agent was malicious; it was that it had not learned "human rules". Humans know that "a 500,000 reimbursement requires the general manager's signature", while the Agent saw "reimbursement process -> check approver -> submit", and it selected a department manager and submitted.

This reminds me of a saying:The higher the efficiency, the heavier the responsibility.

The tool gave us fast hooves, but the horse tamer did not keep up.

Three, horses and horse tamers: two roles in the AI era

Back to the topic of "AI horse-raising".

I increasingly feel that human society is dividing into two roles:"Horse keepers" and "horse riders"。

Horse keepers—They are the developers, trainers, and operators of AI Agents. They know the Agent's temperament, boundaries, and weaknesses. They feed data to the Agent, set rules, and define boundaries. They are like ancient horse trainers, knowing the habits of each horse.

Horse riders—They are the users of AI Agents. They only need to tell the Agent "where I want to go," and leave the rest to the Agent. They don't care about technical details, only results.

These two roles are indispensable.

But today's problem is:Too many people are "riding horses," too few are "keeping horses."

Enterprises are frantically deploying AI Agents to improve efficiency, but the "domestication" work of Agents—rule setting, boundary management, exception handling, continuous training—lags far behind.

This is like raising a thousand-mile horse, but you never feed it, never brush it, never train it, and still expect it to travel a thousand miles a day.

Impossible.

The EU has already realized this. In the recently issued implementation details of the "AI Act," it explicitly requires enterprises deploying AI Agents to establish an "AI Ethics Compliance Officer" position. This is essentially a "Chief Horse Trainer."

In China, the WeChat Mini Program ecosystem recently announced the opening of AI capabilities to developers—this means that tens of millions of Mini Program developers will become "horse keepers." This is a huge opportunity and also a huge responsibility.

IV. Three Suggestions on AI Agents

Having discussed the phenomena and trends, I want to give you three actionable suggestions.

Suggestion 1: Put a "bridle" on your AI Agent.

If you are using AI Agents to handle business, please make sure to:

First, operations must be auditable. Every Agent operation should be recorded and traceable at any time.

Second, minimize permissions. The data an Agent can access should be less, not more, than that of your ordinary employees.

Third, human fallback. For sensitive operations (deleting data, large approvals, external publishing), there must be a human in the loop.

That case of mistakenly deleting a database ultimately came down to the lack of a "bridle"—the Agent directly obtained root access to the production database.

Suggestion 2: Build "AI stable" infrastructure.

Future enterprises will need three types of new infrastructure:

• AI Operations Platform—Unified management of the operational status, resource consumption, and performance metrics of all AI Agents.

• AI Knowledge Base—Accumulate industry experience, business rules, and historical cases for Agents to learn from and reference.

• AI Security Boundary—Similar to a firewall, but targeting the behavioral boundaries of AI Agents.

These infrastructures are the "stables" for AI Agents. A horse running wild in the field is a beast; domesticated in a stable, it is a tool.

Suggestion 3: Cultivate your "horse trainer" team.

AI Agents are not ready to run just by installing them. They require continuous training, continuous optimization, and continuous correction.

The people you need are not those who "can use AI," but those who "can train AI." There is a world of difference between the two.

People who can use AI: write prompts, let AI do one thing.

People who can train AI: understand the boundaries of AI models, design training data, annotate feedback, and optimize behavioral logic.

The former may only need a week of training. The latter requires 3-6 months of systematic learning.

But the return is enormous—An excellent "horse trainer" can increase a team's AI efficiency by 5-10 times.